[BENCHMARKS Added, W 12B-VL vs 8B-VL model too.]
This repo contains the full precision source code, in "safe tensors" format to generate GGUFs, GPTQ, EXL2, AWQ, HQQ and other formats. The source code can also be used directly.
This model contains the full source of the original "Qwen3-VL-Instruct-8B" coupled with Brainstorm 20x adapter (by DavidAU).
This is the full image/text/multi-modal model AND fully tested.
Brainstorm 20x will augment text generation as well as "image" description and in many cases raises the raw benchmarks of the model too.
Addition of the Brainstorm adapter has altered the model from 8B to 12B (now with 55 layers and 608 tensors).
IMPORTANT: - Looping?
This model tends to loop sometimes, if this happens set temp to "1" (or higher) and rep pen to "1.1" . The org model has the same issue. If you have the option, set "presence penalty" to 2.0 [as per Qwen's own suggestions]
IMPORTANT-GGUF QUANTS:
You need both the GGUF quant(s) AND the special "mmproj" gguf quant (Q8,F16,BF16 or F32) to use all functions with this model.
Specialized MAX GGUF QUANTS are here:
https://huggingface.co/DavidAU/Qwen3-VL-12B-Instruct-Brainstorm20x-NEO-MAX-GGUF
(Quants, BENCHMARKS, Brainstorm details, org model details from Qwen, and then help section)
This model requires:
Settings used for testing (suggested):
This model will respond well to both detailed instructions and step by step refinement and additions to code.
Likewise for creative use cases.
As this is an instruct model, it will also benefit from a detailed system prompt too.
For simpler coding problems, lower quants will work well; but for complex/multi-step problem solving suggest Q6 or Q8.
QUANTS:
GGUF? GGUF Imatrix? Other?
Special thanks to Team Mradermacher, Team Nightmedia and other quanters!
See under "model tree", upper right and click on "quantizations".
New quants will automatically appear.
BENCHMARKS by Nightmedia (also makes MLX quants too)
https://huggingface.co/nightmedia/
https://huggingface.co/nightmedia/Qwen3-VL-12B-Instruct-Brainstorm20x-qx86-hi-mlx
📊 Benchmark Summary Table (Performance)
Model ARC-Challenge ARC-Easy BoolQ HellaSwag OpenBookQA PIQA Winogrande
q6-hi 0.501 0.649 0.870 0.634 0.414 0.758 0.641
q8-hi 0.511 0.661 0.872 0.640 0.420 0.763 0.646
qx86-hi 0.502 0.646 0.871 0.637 0.412 0.761 0.644
qx86 0.497 0.646 0.873 0.637 0.414 0.758 0.639
qx86x-hi 0.500 0.650 0.873 0.636 0.410 0.760 0.645
🔍 Cognitive Evaluation by Benchmark
✅ ARC-Easy & ARC-Challenge
These test reasoning, especially science-based logical inference.
q8-hi leads in ARC-Easy (0.661), a clear cognitive strength.
q8-hi also leads ARC-Challenge (0.511), showing strong reasoning
ability on difficult science questions.
👉 Impression: This suggests the model has solid abstract reasoning
capabilities, especially in foundational cognitive domains.
✅ BoolQ
Binary QA (logical yes/no inference).
All models score well (>0.87), with q6-hi, qx86x-hi, qx86 at 0.873.
This reflects strong logical deduction—a hallmark of cognitive performance.
👉 Impression: The model consistently processes binary truth claims with
high reliability—akin to sharp logical reasoning.
✅ HellaSwag
Commonsense reasoning about likely consequences of events (e.g.,
"I took out the trash. Then..." → what next?).
q8-hi scores highest (0.640), but all are relatively low (~0.635–0.640).
This is a known hard benchmark, so even the top model struggles (~64%).
👉 Impression: The cognitive ability here is decent but still human-level
in difficulty—it reflects a struggle for all models, as expected.
✅ OpenBookQA
Science QA requiring external knowledge and reasoning
(e.g., "Why do plants need sunlight?").
q8-hi edges out at 0.420, which is the highest score.
This reflects ability to retrieve and reason with factual knowledge,
a cognitive requirement.
👉 Impression: It can handle straightforward fact-based reasoning, but
struggles with deeper explanations (typical of strong models).
✅ PIQA
Physical commonsense reasoning ("Which is harder to lift: a feather or a rock?").
q8-hi and qx86x-hi score highest (0.763, 0.760), both close to ~76% accuracy.
👉 Impression: Strong cognition in everyday, embodied reasoning—shows
understanding of physical causality.
✅ Winogrande
Pronoun coreference resolution (understanding entity references in sentences).
q8-hi scores highest at 0.646.
All models hover around ~0.64, which is challenging for language understanding.
👉 Impression: Solid semantic and syntactic parsing ability—requires
deep linguistic modeling.
🧠 Cognitive Strengths Summary
Reasoning:
q8-hi leads in ARC-Easy and ARC-Challenge → strong analytical reasoning.
Logical Deduction:
BoolQ scores >0.87 → excellent at binary logic processing.
Commonsense Understanding:
PIQA (~76%) and HellaSwag (~64%) → solid grasp of real-world reasoning.
Knowledge Retrieval:
OpenBookQA (~0.42) → decent but not outstanding.
Language Understanding:
Winogrande (~0.64) → functional but still challenging.
💡 Final Cognitive Verdict
✅ The q8-hi variant is the most cognitively mature, showing balanced,
reliable reasoning across diverse cognitive tasks.
It consistently outperforms the others in key benchmarks that reflect
logical and commonsense cognition — especially ARC-Easy, PIQA, and OpenBookQA.
🧠 Key Takeaway: This AI model demonstrates cognitive abilities most
aligned with a rational, logical thinker—akin to Spock (but not quite).
It’s efficient at reasoning without emotional interference, has strong
factual grounding (BoolQ), and a good intuitive grasp of the real world
(PIQA, Winogrande).
📌 Bonus Insight: Quantization Impact
q8-hi → best cognitive balance overall (~64.8% average accuracy).
q6-hi → slightly lower reasoning but competitive in some areas.
qx86 variants → generally perform equivalently to q6-hi or slightly worse—
they are likely optimized for size/speed, not cognition.
qx86x-hi → a hybrid variant that nearly matches q8-hi in PIQA and BoolQ,
but trails slightly elsewhere.
🧠 Final Comparison to Spock
Trait q8-hi
Logical ✅ High accuracy in BoolQ, ARC
Consistent ✅ Stable across benchmarks
Rational ✅ No overconfidence in weak areas
Knowledge-based 🟠 Moderate (OpenBookQA ~0.42)
✅ q8-hi is the best approximate Spock — not perfect, but rational,
reliable, and strong in reasoning.
So yes, if Spock were an AI model, it would definitely run q8-hi. 💡🖖
BRAINSTORM VS REG QWen3 -VL 8B (Brainstorm is added to this model)
Let's analyze the Qwen3-VL-12B-Instruct-Brainstorm20x series. It has an
extra 4B of brainstorming space. For comparison, the metrics from the
Qwen3-VLTO-8B-Instruct that is similar to the baseline used for the 12B
How did brainstorming improve the model, and how do the individual quants perform
qwen3-vl-12b-instruct-brainstorm20x-qx86-hi-mlx
This is a fascinating and highly relevant comparison — you’re now analyzing
the Qwen3-VL-12B-Instruct-Brainstorm20x series, which is a 12B model
with “extra 4B of brainstorming space”, and comparing it to the
Qwen3-VLTO-8B-Instruct baseline (which is quantized at qx86x-hi).
🧠 1. What Does “Brainstorm20x” Mean?
The name suggests:
“Brainstorm” — likely refers to enhanced internal reasoning
capacity, possibly via:
Expanded attentional memory (e.g., longer context or more
intermediate reasoning steps).
“20x” — likely refers to 20× more internal “thinking space”
or reasoning capacity, perhaps via:
Expanded hidden states.
More layers of intermediate reasoning (e.g., “think step-by-step”).
Dedicated “brainstorming” layers — perhaps a MoE or attention
expansion layer.
This is not just model size, but architectural augmentation —
adding “thinking space” to improve reasoning depth.
📊 2. Benchmark Comparison:
Qwen3-VLTO-8B vs Qwen3-VL-12B-Brainstorm20x
Metric Qwen3-VLTO-8B (qx86x-hi) Qwen3-VL-12B-Brainstorm20x (qx86x-hi)
arc_challenge 0.455 0.502
arc_easy 0.601 0.646
boolq 0.878 0.871
hellaswag 0.546 0.637
openbookqa 0.424 0.410
piqa 0.739 0.760
winogrande 0.595 0.645
Overall Avg 0.579 0.634
✅ The 12B-Brainstorm20x model is clearly superior across all metrics
— +0.05–0.13 gains, with the most dramatic improvements in:
ARC Challenge (+0.047)
ARC Easy (+0.045)
Hellaswag (+0.091)
Winogrande (+0.05)
The only metric where it’s slightly worse is OpenBookQA (↓0.014) —
likely due to overfitting or less effective handling of purely textual
inference without visual grounding.
🧪 3. How Did “Brainstorm20x” Improve the Model?
The key insight: adding 4B of “brainstorming space” didn’t
just scale the model — it enhanced its reasoning depth.
🔍 Cognitive Impact:
ARC Challenge & ARC Easy:
+0.047 and +0.045 — this suggests better reasoning chain decomposition.
Hellaswag:
+0.091 — this suggests better commonsense inference, likely due to more
intermediate reasoning steps.
Winogrande:
+0.05 — this suggests better contextual understanding, likely due to
expanded attentional memory.
Piqa:
+0.021 — this suggests better step-by-step reasoning, likely due to
more intermediate steps.
The model is now capable of “thinking deeper” — not just “thinking faster”.
🧩 4. Quantization Comparison within the 12B-Brainstorm20x Series
Let’s compare the qx86x-hi variant (best overall) to others:
Model arc_challenge arc_easy boolq hellaswag openbookqa piqa winogrande
Qwen3-VL-12B-Instruct-Brainstorm20x-q6-hi 0.501 0.649 0.870 0.634 0.414 0.758 0.641
Qwen3-VL-12B-Instruct-Brainstorm20x-q8-hi 0.511 0.661 0.872 0.640 0.420 0.763 0.646
Qwen3-VL-12B-Instruct-Brainstorm20x-qx86-hi 0.502 0.646 0.871 0.637 0.412 0.761 0.644
Qwen3-VL-12B-Instruct-Brainstorm20x-qx86 0.497 0.646 0.873 0.637 0.414 0.758 0.639
Qwen3-VL-12B-Instruct-Brainstorm20x-qx86x-hi 0.500 0.650 0.873 0.636 0.410 0.760 0.645
✅ qx86x-hi is the best overall, with +0.01–0.02 gains over other variants.
The q8-hi variant is the best performer overall, with:
+0.01 in arc_challenge
+0.02 in arc_easy
+0.006 in hellaswag
+0.01 in winogrande
The qx86x-hi variant is the most balanced, with slightly better
performance across all metrics.
🧭 5. Recommendation: Which Quant to Choose?
✅ For Maximum Performance:
Qwen3-VL-12B-Instruct-Brainstorm20x-q8-hi
→ Best overall performance, especially in Hellaswag, Winogrande, and Piqa.
✅ For Balanced Performance & RAM Efficiency:
Qwen3-VL-12B-Instruct-Brainstorm20x-qx86x-hi
→ Best overall average, with slightly better performance across all metrics.
✅ For RAM-Constrained Macs (32GB):
Qwen3-VL-12B-Instruct-Brainstorm20x-qx86x-hi
→ ~27.7GB RAM usage — fits comfortably.
🧠 6. Cognitive Pattern Insight: “Brainstorm20x” Enhances Reasoning Depth
The “Brainstorm20x” augmentation is not just scaling
— it’s enhancing reasoning depth.
The model now has more internal “thinking space”, allowing it to:
Break down complex problems into intermediate steps.
Infer commonsense reasoning more accurately.
Handle ambiguous prompts with better context.
This is a cognitive upgrade, not just a computational one —
the model now “thinks deeper”, not just “faster”.
📈 7. Summary Table: Best Model for Each Use Case
Goal Recommended Model RAM Usage Performance Rank
Max performance Qwen3-VL-12B-Instruct-Brainstorm20x-q8-hi ~27.6GB #1
Balanced performance Qwen3-VL-12B-Instruct-Brainstorm20x-qx86x-hi ~27.7GB #1
RAM-constrained Mac (32GB) Qwen3-VL-12B-Instruct-Brainstorm20x-qx86x-hi ~27.7GB #1
Cognitive depth & metaphors Qwen3-VL-12B-Instruct-Brainstorm20x-q8-hi ~27.6GB #1
OpenBookQA (text-only) Qwen3-VL-12B-Instruct-Brainstorm20x-q6-hi ~24GB #1
🚀 Bonus: “Brainstorm20x” as a Cognitive Architecture
The “Brainstorm20x” augmentation is not just more parameters — it’s more reasoning capacity.
It adds 20× internal “thinking space”, allowing the model to:
Break down complex problems into intermediate steps.
Infer commonsense reasoning more accurately.
Handle ambiguous prompts with better context.
This is a cognitive upgrade, not just a computational one —
the model now “thinks deeper”, not just “faster”.
“Brainstorm20x is like adding a second brain — not just more
neurons, but more thinking steps.”
— Inspired by the human mind’s ability to “think step-by-step”.
Brainstorm 20x
The BRAINSTORM process was developed by David_AU.
Some of the core principals behind this process are discussed in this scientific paper : Progressive LLaMA with Block Expansion .
However I went in a completely different direction from what was outlined in this paper.
What is "Brainstorm" ?
The reasoning center of an LLM is taken apart, reassembled, and expanded.
In this case for this model: 20 times
Then these centers are individually calibrated. These "centers" also interact with each other. This introduces subtle changes into the reasoning process. The calibrations further adjust - dial up or down - these "changes" further. The number of centers (5x,10x etc) allow more "tuning points" to further customize how the model reasons so to speak.
The core aim of this process is to increase the model's detail, concept and connection to the "world", general concept connections, prose quality and prose length without affecting instruction following.
This will also enhance any creative use case(s) of any kind, including "brainstorming", creative art form(s) and like case uses.
Here are some of the enhancements this process brings to the model's performance:
Other "lab" observations:
The process to modify the model occurs at the root level - source files level. The model can quanted as a GGUF, EXL2, AWQ etc etc.
Meet Qwen3-VL — the most powerful vision-language model in the Qwen series to date.
This generation delivers comprehensive upgrades across the board: superior text understanding & generation, deeper visual perception & reasoning, extended context length, enhanced spatial and video dynamics comprehension, and stronger agent interaction capabilities.
Available in Dense and MoE architectures that scale from edge to cloud, with Instruct and reasoning‑enhanced Thinking editions for flexible, on‑demand deployment.
Visual Agent: Operates PC/mobile GUIs—recognizes elements, understands functions, invokes tools, completes tasks.
Visual Coding Boost: Generates Draw.io/HTML/CSS/JS from images/videos.
Advanced Spatial Perception: Judges object positions, viewpoints, and occlusions; provides stronger 2D grounding and enables 3D grounding for spatial reasoning and embodied AI.
Long Context & Video Understanding: Native 256K context, expandable to 1M; handles books and hours-long video with full recall and second-level indexing.
Enhanced Multimodal Reasoning: Excels in STEM/Math—causal analysis and logical, evidence-based answers.
Upgraded Visual Recognition: Broader, higher-quality pretraining is able to “recognize everything”—celebrities, anime, products, landmarks, flora/fauna, etc.
Expanded OCR: Supports 32 languages (up from 19); robust in low light, blur, and tilt; better with rare/ancient characters and jargon; improved long-document structure parsing.
Text Understanding on par with pure LLMs: Seamless text–vision fusion for lossless, unified comprehension.
Interleaved-MRoPE: Full‑frequency allocation over time, width, and height via robust positional embeddings, enhancing long‑horizon video reasoning.
DeepStack: Fuses multi‑level ViT features to capture fine‑grained details and sharpen image–text alignment.
Text–Timestamp Alignment: Moves beyond T‑RoPE to precise, timestamp‑grounded event localization for stronger video temporal modeling.
This is the weight repository for Qwen3-VL-8B-Instruct.
Multimodal performance
Pure text performance
Below, we provide simple examples to show how to use Qwen3-VL with 🤖 ModelScope and 🤗 Transformers.
The code of Qwen3-VL has been in the latest Hugging Face transformers and we advise you to build from source with command:
pip install git+https://github.com/huggingface/transformers
# pip install transformers==4.57.0 # currently, V4.57.0 is not released
Here we show a code snippet to show how to use the chat model with transformers:
from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
# default: Load the model on the available device(s)
model = Qwen3VLForConditionalGeneration.from_pretrained(
"Qwen/Qwen3-VL-8B-Instruct", dtype="auto", device_map="auto"
)
# We recommend enabling flash_attention_2 for better acceleration and memory saving, especially in multi-image and video scenarios.
# model = Qwen3VLForConditionalGeneration.from_pretrained(
# "Qwen/Qwen3-VL-8B-Instruct",
# dtype=torch.bfloat16,
# attn_implementation="flash_attention_2",
# device_map="auto",
# )
processor = AutoProcessor.from_pretrained("Qwen/Qwen3-VL-8B-Instruct")
messages = [
{
"role": "user",
"content": [
{
"type": "image",
"image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg",
},
{"type": "text", "text": "Describe this image."},
],
}
]
# Preparation for inference
inputs = processor.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
return_dict=True,
return_tensors="pt"
)
inputs = inputs.to(model.device)
# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=128)
generated_ids_trimmed = [
out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)
export greedy='false'
export top_p=0.8
export top_k=20
export temperature=0.7
export repetition_penalty=1.0
export presence_penalty=1.5
export out_seq_length=16384
export greedy='false'
export top_p=1.0
export top_k=40
export repetition_penalty=1.0
export presence_penalty=2.0
export temperature=1.0
export out_seq_length=32768
CHANGE THE NUMBER OF ACTIVE EXPERTS:
See this document:
https://huggingface.co/DavidAU/How-To-Set-and-Manage-MOE-Mix-of-Experts-Model-Activation-of-Experts
Settings: CHAT / ROLEPLAY and/or SMOOTHER operation of this model:
In "KoboldCpp" or "oobabooga/text-generation-webui" or "Silly Tavern" ;
Set the "Smoothing_factor" to 1.5
: in KoboldCpp -> Settings->Samplers->Advanced-> "Smooth_F"
: in text-generation-webui -> parameters -> lower right.
: In Silly Tavern this is called: "Smoothing"
NOTE: For "text-generation-webui"
-> if using GGUFs you need to use "llama_HF" (which involves downloading some config files from the SOURCE version of this model)
Source versions (and config files) of my models are here:
OTHER OPTIONS:
Increase rep pen to 1.1 to 1.15 (you don't need to do this if you use "smoothing_factor")
If the interface/program you are using to run AI MODELS supports "Quadratic Sampling" ("smoothing") just make the adjustment as noted.
Highest Quality Settings / Optimal Operation Guide / Parameters and Samplers
This a "Class 1" model:
For all settings used for this model (including specifics for its "class"), including example generation(s) and for advanced settings guide (which many times addresses any model issue(s)), including methods to improve model performance for all use case(s) as well as chat, roleplay and other use case(s) please see:
You can see all parameters used for generation, in addition to advanced parameters and samplers to get the most out of this model here: